Click here to


Are you sure ?

Yes, do it No, cancel

How Many Sample Sizes Are Appropriate for Deep Learning Based Auto Segmentation for Head and Neck Cancer?

F Yingtao , W Hu*, J Wang , S Chen , S Sun , Z Zhang , Fudan University Shanghai Cancer Center, Shanghai


(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 5

Purpose: To investigate the effect of sample sizes on auto segmentation with convolutional neural network for head and neck organ.

Methods: A total of 1,160 patients with head and neck cancer were enrolled in the study. Physicians had delineated 14 organs including brain stem, spinal cord, left and right eyes, left and right lenses, left and right optic nerves, left and right temporal lobes, left and right parotids, larynx and body. The population is randomly divided into evaluation dataset and training dataset. The evaluation dataset has 200 patients. Four different training dataset, including 200, 400, 600, 800 patients, were randomly selected from the remaining 960 patients. Four auto segmentation models were developed (m200, m400, m600 and m800). All models use same structure and training hype-parameter. The evaluation dataset were used to evaluate the quality of these auto segmentation models. We calculated the dice index (DSC) and compared the value between different auto segmentation models.

Results: M800 have the best performance in all models in which all organs’ DSC were over 0.65 and nine of them were over 0.75. The average DSC of all organs in m200, m400, m600 and m800 are 0.7579, 0.7666, 0.7669 and 0.7673. Organs could be divided into two categories according to difficulty of auto segmentation. Meanwhile, the improvement of the performance did not have obviously pattern.

Conclusion: The sample size has influence on the performance of deep learning-based auto segmentation. Difficult organ can get more improvement by increasing training dataset.


Image Analysis


IM/TH- image segmentation: CT

Contact Email